Post

Very proud to share this new work led by the amazing Marlou Rasenberg (not on here), with Joel Fischer and me doi.org/10.5281/zeno... We use #EMCA methods to study empirically how LLM users shoulder the burden of solving sequential trouble and end up restructuring their own aims & activities

We identify a set of practices we call 'pre-emptive prompting', "through which users creatively circumvent the system’s lack of action recognition and action projection". Users stumble upon these through trial and error, in exchanges like this:

While much work in so-called 'prompt engineering' obsesses over the perfect prompt (Andreessen's monster prompt comes to mind), we find that real-life prompting practices are sequentially organized, feature repurposed patterns of human interaction, and result in users adjusting expectations & aims

We build on work by @abeba.blacksky.app @marekmcgann.bsky.social @5tuartreeves.bsky.social @lizstokoe.bsky.social @emilymbender.bsky.social and classics like Boden, Colby, Terasaki &c More later — for now, here's the preprint doi.org/10.5281/zeno... #prompting #emca

Super interesting (and appreciate as always that it's accessible to someone like me)! It seems like hidden labor is not uncommon in sw products, but ime because it's better masked in llms, coworkers can fail to realize they should have kept laboring a lot longer before sharing their output.

I can just see no logical inference except that a tool that mimics thought in language without being it will always and inevitably increase/generate workload for any mind participating, and the more convincing the fakery the greater the workload increase.

You mean no logical inference on the part of the LLM? That tracks

I think so. It has reasoning. But also in the system as a whole, we respond automatically to beings with minds, so it's always intoducing a massive source of error that, at best, we expend great cognive resource and effort to parse/resist/put aside. Just can't see how that could ever aid efficiency.